fix: cli duration outputs

Author: Sean Roberts

src/adapters/base/acp-adapter.ts

@@ -112,8 +112,6 @@ interface AcpState {
   activeToolCalls: Map<string, { title: string; kind?: string }>;
   /** Cumulative session cost from usage_update events. */
   totalCostUsd?: number;
-  /** Duration of the prompt() call only (excludes ACP handshake and process lifecycle). */
-  promptDurationMs?: number;
 }
 
 // ---------------------------------------------------------------------------
@@ -289,13 +287,15 @@ export function createAcpBasedAdapter(spec: AcpAdapterSpec): AgentAdapter {
           input.onRawLine?.(line);
         }
 
-        // 14. Send prompt and wait for completion — time just the agent work
-        const promptStart = Date.now();
+        // 14. Send prompt and wait for completion. Signal "agent is now
+        // doing real work" — for ACP adapters the initialize+newSession
+        // handshake can take 10–20s (kiro-cli AWS auth, etc.) and shouldn't
+        // count as `running` in the live UI.
+        input.onAgentReady?.();
         const promptResult = await connection.prompt({
           sessionId: sessionResult.sessionId,
           prompt: [{ type: "text", text: input.prompt }],
         });
-        state.promptDurationMs = Date.now() - promptStart;
         {
           const line = JSON.stringify({ type: "prompt_result", ...promptResult });
           rawOutput?.push(line);
@@ -391,7 +391,7 @@ export function createAcpBasedAdapter(spec: AcpAdapterSpec): AgentAdapter {
       const metadata: AgentMetadata = {
         startTime: startTime.toISOString(),
         endTime: endTime.toISOString(),
-        durationMs: state.promptDurationMs ?? endTime.getTime() - startTime.getTime(),
+        durationMs: endTime.getTime() - startTime.getTime(),
         exitCode,
         tokenUsage: state.tokenUsage,
         totalCostUsd: state.totalCostUsd,

src/report-ui/src/scripts/render.ts

@@ -767,9 +767,13 @@ function renderWaterfall(si: SparseIndex, totalDurationMs: number): string {
   const hasTimingData = interactions.some((ix) => ix.startMs !== null);
   if (!hasTimingData || interactions.length === 0) return "";
 
-  // Prefer the agent's full run duration so the timeline matches the row's reported duration.
-  // Fall back to the interaction window if the entry's duration is missing.
-  const wallClockMs = totalDurationMs > 0 ? totalDurationMs : si.stats.wallClockMs || computeWallClock(interactions);
+  // Interaction startMs values are measured from process spawn (wall clock),
+  // so the chart axis must use wall-clock too. For ACP-based adapters
+  // entry.durationMs is the prompt() time only (excludes handshake/shutdown)
+  // and is smaller than the interactions' timeline — using it would push bars
+  // off the right edge. Prefer sparse-index wallClockMs, fall back to entry
+  // duration for adapters without stats, then to the interaction window.
+  const wallClockMs = si.stats.wallClockMs || (totalDurationMs > 0 ? totalDurationMs : computeWallClock(interactions));
   if (wallClockMs <= 0) return "";
 
   const ticks = computeTickMarks(wallClockMs);

src/runner/runner.ts

@@ -251,15 +251,22 @@ export async function run(options: RunOptions = {}): Promise<RunOutput> {
 
   const updateStatus = (index: number, status: JobStatus, durationMs?: number) => {
     const patch: Partial<JobState> = { status, durationMs };
-    // Stamp the start time on the first transition into "running" so the
-    // live UI can tick an elapsed-duration counter.
-    if (status === "running" && jobStates[index].runStartedAt === undefined) {
+    // Stamp the start time on the first transition into "starting"/"running"
+    // so the live UI can tick an elapsed-duration counter that includes the
+    // adapter's own startup (CLI cold start, ACP handshake, etc.).
+    if ((status === "starting" || status === "running") && jobStates[index].runStartedAt === undefined) {
       patch.runStartedAt = Date.now();
     }
     jobStates[index] = { ...jobStates[index], ...patch };
     logger.onJobUpdate?.(jobStates, jobMeta);
   };
 
+  const promoteToRunning = (index: number) => {
+    if (jobStates[index].status === "starting") {
+      updateStatus(index, "running");
+    }
+  };
+
   const setTeardown = (index: number, inTeardown: boolean) => {
     if (Boolean(jobStates[index].inTeardown) === inTeardown) return;
     jobStates[index] = { ...jobStates[index], inTeardown };
@@ -443,6 +450,7 @@ export async function run(options: RunOptions = {}): Promise<RunOutput> {
         logger,
         updateStatus,
         updateTokens,
+        promoteToRunning,
         resolvedSkillMap,
         options.registerCleanup,
         options.debug ?? false,
@@ -492,6 +500,7 @@ async function executeJob(
   logger: Logger,
   updateStatus: (index: number, status: JobStatus, durationMs?: number) => void,
   updateTokens: (index: number, tokens: number, final?: boolean) => void,
+  promoteToRunning: (index: number) => void,
   resolvedSkillMap: Map<string, ResolvedSkill>,
   registerCleanup?: (fn: () => void) => void,
   debug?: boolean,
@@ -618,7 +627,7 @@ async function executeJob(
   }
 
   try {
-    updateStatus(index, "running");
+    updateStatus(index, "starting");
     logger.verbose?.(`[${label}] Executing agent...`);
 
     // Merge top-level + per-agent + per-scenario skills, deduplicate by source
@@ -648,6 +657,8 @@ async function executeJob(
         : axisConfig.mcp_servers,
       resolvedSkills: agentSkills.length > 0 ? agentSkills : undefined,
       onTokenProgress: (tokens) => {
+        // First token from the agent → it's past startup, into real work.
+        promoteToRunning(index);
         updateTokens(index, tokens);
         // Per-scenario token limit
         if (jobLimits?.tokenLimit && tokens >= jobLimits.tokenLimit) {
@@ -656,6 +667,7 @@ async function executeJob(
         // Overall token limit (checks cumulative across all jobs)
         checkOverallTokenLimit?.();
       },
+      onAgentReady: () => promoteToRunning(index),
       ...(jobLimits?.timeoutMs ? { timeoutMs: jobLimits.timeoutMs } : {}),
       ...(jobAbortController ? { signal: jobAbortController.signal } : {}),
       debug,

src/types/agent.ts

@@ -86,6 +86,14 @@ export interface AgentInput {
    * below the true count so the UI never has to reverse.
    */
   onTokenProgress?: (estimatedTokens: number) => void;
+  /**
+   * Adapters call this when the agent has finished its own startup work
+   * (CLI cold start, ACP handshake, session bootstrap) and is actually
+   * processing the prompt. The runner uses it to flip job status from
+   * `starting` to `running`. Adapters with negligible startup may skip
+   * calling it — the first `onTokenProgress` will auto-promote.
+   */
+  onAgentReady?: () => void;
   /** Override the adapter's default timeout (in ms). Set by the runner from resolved scenario limits. */
   timeoutMs?: number;
   /** Abort signal. When fired, the adapter kills the child process with SIGTERM → SIGKILL. */

src/types/output.ts

@@ -64,7 +64,7 @@ export interface RunSummary {
   skipped?: number;
 }
 
-export type JobStatus = "pending" | "setup" | "running" | "teardown" | "done" | "failed" | "scoring";
+export type JobStatus = "pending" | "setup" | "starting" | "running" | "teardown" | "done" | "failed" | "scoring";
 
 export interface JobState {
   scenarioKey: string;
@@ -87,9 +87,10 @@ export interface JobState {
    */
   tokensFinal?: boolean;
   /**
-   * Wall-clock ms-epoch when the agent transitioned to `running`. Used by the
-   * live UI to tick an elapsed-duration counter before the job finishes (once
-   * finished, `durationMs` takes over as the authoritative value).
+   * Wall-clock ms-epoch when the agent transitioned out of `pending` into
+   * `starting`/`running`. Used by the live UI to tick an elapsed-duration
+   * counter before the job finishes (once finished, `durationMs` takes over
+   * as the authoritative value).
    */
   runStartedAt?: number;
   /**

src/ui/LiveStatus.tsx

@@ -12,7 +12,13 @@ interface LiveStatusProps {
 /** A job is "active" when it's mid-flight — counts toward the live scenario list. */
 function isActive(job: JobState): boolean {
   if (job.inTeardown) return true;
-  return job.status === "setup" || job.status === "running" || job.status === "teardown" || job.status === "scoring";
+  return (
+    job.status === "setup" ||
+    job.status === "starting" ||
+    job.status === "running" ||
+    job.status === "teardown" ||
+    job.status === "scoring"
+  );
 }
 
 export function LiveStatus({ jobs, skippedCount = 0 }: LiveStatusProps) {
@@ -109,15 +115,15 @@ function AgentRow({ job }: { job: JobState }) {
       ? "green"
       : job.status === "failed"
         ? "red"
-        : job.status === "running" || job.status === "scoring"
+        : job.status === "starting" || job.status === "running" || job.status === "scoring"
           ? "yellow"
           : undefined;
 
   // The timer and token counter are both visible in every state where they
-  // have a value. The timer shows live-elapsed during running/scoring and
-  // the final `durationMs` afterwards. The token counter keeps animating
+  // have a value. The timer shows live-elapsed during starting/running/scoring
+  // and the final `durationMs` afterwards. The token counter keeps animating
   // until it catches up to the final real total even after the job is done.
-  const active = job.status === "running" || job.status === "scoring";
+  const active = job.status === "starting" || job.status === "running" || job.status === "scoring";
   const hasTime = job.runStartedAt !== undefined || job.durationMs !== undefined;
   const hasTokens = (job.liveTokens ?? 0) > 0;
 

src/ui/format.ts

@@ -45,6 +45,7 @@ const CRITERION_MEDIUM = 4;
 export const STATUS_ICONS: Record<string, string> = {
   pending: "○",
   setup: "◌",
+  starting: "◔",
   running: "●",
   teardown: "◌",
   done: "✓",
@@ -55,6 +56,7 @@ export const STATUS_ICONS: Record<string, string> = {
 export const STATUS_LABELS: Record<string, string> = {
   pending: "pending",
   setup: "setup",
+  starting: "starting",
   running: "running",
   teardown: "teardown",
   done: "done",